FLOWSHOP SCHEDULES WITH SEQUENCE DEPENDENT SETUP TIMES

Transcription

1 Journal of the Operations Research Society of Japan Vo!. 29, No. 3, Septeber The Operations Research Society of Japan FLOWSHOP SCHEDULES WITH SEQUENCE DEPENDENT SETUP TIMES Jatinder N. D. Gupta Ball State University (Received June 24,1985: Final April 30, 1986) Abstract The flowshop scheduling proble with sequence dependent setup ties is considered and shown to be NP-coplete. A traveling salesan proble forulation is proposed for the case where jobs are processed continuously through the shop. These results are used to describe an approxiate algorith for the case where liited or infinite interediate storage space is available to hold partially copleted jobs. The effectiveness of the proposed approxiate approach is discussed and soe epirical results are reported. 1. Introduction Consider a static flowshop scheduling proble where n jobs are to be processed on M achines in the sae technological order. In forulating this proble, it is usually assued that setup ties are sequence independent and are included in the processing ties. However, there are several practical situations, where setup ties of a job are separable and depend on the iediately preceding job. As an illustration, consider the scheduling proble in the group technology environent. For each faily of parts, a large setup tie is required to initiate the processing of faily parts after which sall setup tie is required and is dependent on the sequence of jobs preceding a particular faily part (job) being processed [10]. When setup ties are separable and sequence independent, and infinite interediate space is available to hold jobs, Johnson's algorith [13] can be odified to generate a inial ake-span schedule in O(nlognJ coputational steps [20] for the two achine case. Further, if the setup ties are separable and sequence dependent on only one of the two achines, a dynaic prograing approach can be used to iniize the akespan for probles involving about 15 jobs [2, 4]. Heuristic algoriths for iniizing the akespan for the two 206

2 Sequence-Dependent F1owshop Scheduling 207 achine case are also available [10, 11]. If the setup ties are separable and sequence dependent on all achines, atheatical prograing and iplicit enljideration approaches that are based on its graph-theoretic representation can be odified to solve the proble [6,9, 17, 18]. However, the coputational requireents for the dynaic prograing and the branch and bound algoriths are quite excessive even for probles of oderate size. This paper discusses the flowshop seheduling proble with sequence dependent setup ties and shows that the proble is NP-coplete. A traveling salesan proble forulation is suggested for the case where the processing of jobs through the shop is continuous. Based on the traveling salesan forulation of the continuous processing flowshop, an approxiation solution algorith is proposed for the flowshop proble with liited or infinite interediate storage space in the presence of sequence dependent setup ties. The effectiveness of the proposed approxiate algorith is epirically evaluated by solving several flowshop probles and the results of this coparison are discussed. 2. Proble Coplexity and Forulation The following theore shows that the flowshop proble with sequence dependent set up ties and any easure of perforance which is a function of the processing and setup ties is NP-co~)lete. Theore 1. ties is NP-coplete. Proof: The flowshop scheduling proble with sequence dependent setup The proof follows by restriction [5]. Let the setup and proc essing ties of all jobs on any cobination of (M-l) achines be zero. Then, this is a classical single facility proble with sequence dependent setup ties and has been shown to be NP-coplete [5], iplying that the proble under consideration is also NP-coplete. Hence the proof of theore 1. The results of theore 1 above are independent of the nuber of achines and the interediate storage space available to hold partially cop leted jobs. Therefore, unlike the cases where setup ties are either sequence independent or can be ignored, the proble reains NP-coplete even for the two-achine case. Further, the three cases, Viz: continuous processing, infinite interediate storage available, and the finite! storage space available cases are all NP-coplete. In fact, for the two achine case, perutation schedules are not necessarily optial [10]. To siplify the proble, only perutation

3 208 J. N. D. Gupta schedules are considered in this paper. This assuption does not affect the global optiality of the continuous processing case, as will be shown shortly. In addition, results of theore 1 above are independent of the easure of perforance used to evaluate schedules. This iplies that flowshop proble with sequence dependent setup ties reains NP-coplete for any easure of perforance involving setup ties even for the two achine case, with or without the assuption of perutation schedules. Let a (i,j) be the setup tie of job j at achine if job i innnediately precedes it, where i=o if no job precedes job j. Further, let t(i,) be the processing tie of job i on achine. Nuber the achines such that jobs are processed on achine 1 first, achine 2 second and the achine M last. With this re indexing of achines, consider a partial schedule Pab fored by concatenating job b to an already known partial schedule Pa. Then the copletion tie of Pab at achine, T(Pab,), is given by the following expression: (1) where T(Pab,) = ax[t(pab,-l); T(Pa,) + a (a,b)] + t(b,) T(4),) = T(Pab,O)=O for all P and. Let the easure of pe!rforance of partial schedule P be f(p) and h(p,a) represents the addition to the easure of perforance of concatenating job a to p. That is: (2) f(pa) f(p) + h(p,a) where f(4))=o. The easure of perforance in (2) above includes several well known optiization criteria for the flowshop scheduling proble. The one used in this paper is the iniization of weighted su of copletion ties on all achines. Let the weight assigned to achine be w. Then, equation (2) becoes: f(pa) M f(p) + L W [T(pa,) - T(P,)] =1 If w M =1 and w=o for =1, 2,.'., (~1), then the above criterion reduces to the iniization of akespan and h(p,a) in equation (2) above is given by: h(p,a) = (T(Pa,M) - T(P,M» The flowshop scheduling proble with sequence dependent setup ties can now be defined as one of iniizing f(pa) where P ranges over all possible perutations of n-l jobs not including job a and job a ranges over all possible values fro 1 through n and that the processing and interediate storage restrictions are satisfied at all achines.

4 Sequence-Dependent Flowshop Scheduling Continuous Processing Flowshop When there is no interediate storage space available to hold partially copleted jobs, the flowshop proble ay be forulated in two ways: (1) a job waits at the first achine before starting its processing such that it can be continuously processed by all achines [8, 17], or (2) a partially copleted job stays at a particular achine until the next one becoes available [14]. In this paper, the continuous case where jobs are delayed before the start of their first operation is considered. nuber of feasible schedules to n! for this case. Theore 2. For a continuous processing flowshop, only perutation schedules are feasible. Proof: The following theore restricts the Consider two consecutive achines 1 and 2 where jobs are processed first on achine 1 and then on achine 2 Assue that job a is processed before job b on achine 1 and that job b is processed before job a on achine 2. Then, since all jobs follow the sae technological processing order, it follows that job a will have to wait at least t(b, 1 ) + a (a,b) + 1 t(b, 2 ) on achine 2 before starting its processing on achine 2 after being copleted on achine 1 Since the setup ties and processing ties are all positive, this waiting tie cannot be zero. Therefore, in the schedule considered, job a cannot be processed continuously on all achines. Since 1, 2, a, and b are all arbitrary, it follows that a feasible schedule has the sae sequence of jobs on all achines. Hence the proof of theore 2. Theore 2 closely follows that in [8]. Since Panwalkar et. al. [15] and Szwarc [19] question the proof for t(a,)=o for soe a and soe rn, coent.s are in order here. First, flowshop scheduling proble can be forulated in two ways, one where each job is routed through each achine even though the jobs do not necessarily require processing by a specific achine and the other where jobs ay skip the achines [1]. In attepting to forulate the flowshop proble, it is usually assued that there is no skipping of jobs. This is equivalent to saying that transportation links exist only between consecutive achines. Second, the flowshop is independent of the scale paraeter iplying that if the paraeters A and B are so chosen that t(a,) = A + Bp(a,) for all a and such that p(a,) is non-negative for all a and but can be zero, the optial schedule for the original proble reains the sae as for the new proble with p(a,) as its processing ties. Therefore, for the forulation of the flowshop that requires routing of jobs through each of the M achines, the results of theore 1 are applicable for t(a,) ~ 0 and a (a,b) ~ 0;

5 210 J. N. D. Gupta ~ a ~ nand 1 ~ ~ M. Details of this arguent are described in [12]. Since there is no interediate storage space to hold partially copleted jobs, jobs ust wait before starting their processing at the first achine. Let the delay in the start of job b after copleting job a at achine 1 such that job b is continuously processed on all M achines be denoted by D(a,b). The tie interval denoted by D(a,b) assues that jobs are processed as early as possible. With this definition of D(a,b), the following theore 3 establishes its value as a function of processing and setup tie of jobs a and b alone. Theore 3. For the continuous processing flowshop where jobs are processed as soon as possible: (3) D(a,b) = ax { ax [ f t(a,5) - ~lt(b,5) + a(a,b»); a 1 (a,b) } 2~~M 5=2 5=1 that: (4) Proof: By definition of continuous processing on all achines, it follows T(Pa,) T(pa,l) + L t(a,5) 5=2 (5) T(pab,l) + L t(b,5) 5=2 Also, fro equation (1), the condition of continuous processing is valid if and only if: T(Pab,rnrl) - T(Pa,) - a (a,b) ~ 0 for all. Siplifying the above with equations (4) and (5) above, results in: -1 (6) L t(a,5) - L t(b,5) + a (a,b) 5=2 5=1 for all. However, D(a,b) is at least as large as a (a,b), the setup tie of job b on 1 achine 1. Therefore, (3) follows fro (6), the above requireent and the fact that jobs are processed as early as possible. Hence the proof of theore 3. Results of theore 3 above show that the initial delay in the start of job b after processing job a is a function the processing ties of jobs a and b alone. This iplies that the proble can be odeled as a traveling salesan proble. To do so, define:

6 Sequence-Dependent Howshop Scheduling 211 a(a,b) (7) a(a,a) M [D(a,b) + t(b,i)] Iw =l fora=o, 1, b=l, 2,..., nand..., n. a(b,o) M I w [ I t (b, s) ] =2 s=2 With the above distances, the following theore can be stated: Theore 4. The schedule S=(a 1, a 2,, an) is optial to the continuous processing flowshop scheduling proble if and only if the tour (0, ai' a,, 2 an' 0) is optial to the traveling salesan proble with distance atrix described by Equation (7) above. by: Proof: The length, L, of the tour (0, ai' a 2,., an' 0) is given n-l L = d(0,a 1 ) + I a(a" a i + 1 ) + a(a n, 0) i";l ~ Siplifying the above expression with Equation (7) yields: (8) L = w 1 A + I W [A+ I t (a, s)] =2 s=2 n where: M n n-l (9) A D(O, a 1 ) + I t(a" 1) + I D(a" a i + 1 ) i=l ~ i=l ~ Now a closer look at (9) reveals that A=T(S,I), the copletion tie of schedule S at achine 1. Therefore, recursive use of Equations (1) and (4) with Equations (8) and (9) results in: (10 ) L M I W [T(S,)] =1 Thus, the total length of the tour (0, ai' a,..., an' 0) in Equation (10) 2 equals the weighted su of copletion ties on all achines for schedule S = (ai' a,..., an). Therefore, an optial schedule for the flowshop 2 proble is an optial tour to the traveling salesan proble (and vice versa) provided the traveling salesan is stationed at city 0 and returns to city 0 after visiting all the n cities. Hence the proof of theore 4. While the theore above showed that the proble is NP-coplete, the traveling salesan forulation of the proble does have soe advantages. First, there are several efficient approxiate solution procedures available

7 212 J. N. D. Gupta to solve the traveling salesan proble (see Parker and Reardon [16] for a recent review), thus providing several approxiate solution procedures for the solution of the continuous processing flowshop scheduling proble with sequence dependent setup ties. Second, worst case analysis of approxiate solution procedures for the traveling salesan proble is soeties available indicating that perforance guarantee for solution procedures for the flowshop proble under consideration is possible. And finally, probabilistic analysis for the perforance of several approxiate algoriths for the traveling salesan proble are now available which can be used for the flowshop proble under consideration. 4. Flowshop with Liited or Infinite Interediate Storage When there is liited or infinite capacity available to hold partially copleted jobs, the results of Theore 2 no longer hold and non-perutation schedules are feasible. In fact, even for the two achine case, perutation schedules are not necessarily optial even for the akespan criterion [10]. Therefore, the no-passing assuption (that only perutation schedules are considered) is active even for the two achine case. The solution of this proble can be approached by iplicit enueration techniques, like the branch and bound procedures [9, 18]. However, these iplicit enueration techniques do not provide a practical approach to solve the proble because of their exponential nature of the coputational effort. Realizing this, ~he traveling salesan approach to the continuous processing flowshop is adopted to generate approxiate schedules for this case as well. This is done by finding an approxiate solution to the traveling salesan proble using a suitable approxiate algorith. This is especially relevant ln Vlew of the fact that the schedules for the two cases are not that far apart fro each other without setup ties [3]. Therefore, the approxiation should work quite well. The proposed approxiate algorith for the liited or infinite storage space flowshop, therefore, ay be described as follows: 1: Proble Forulation: Forulate the traveling salesan proble by using the distances given by Equation (7). 2: Approxiate Solution: Solve the corresponding traveling salesan proble in step 1 using an appropriate approxiate procedure like the "closest city" heuristic. Accept the resultant schedule as an approxiate solution to the flowshop proble.

8 Sequence-Dependent Flowshop Scheduling 213 The appropriate approxiate algorith for the traveling salesan proble in step 2 above can be selected fro several available approxiate solution procedures on the basis of the their coputational effort and the relative accuracy of the solution obtained through their use. (See Parker and Reardon [16] for a recent review of the traveling salesan proble solution procecedures). To easure the effectiveness of the proposed approxiation, let: opt (continuous) - opt (infinite) opt (infinite) Approx(continuous) - opt (continuous) opt (continuous) Approx(infinite) - opt (infinite) e 3 = opt (infini te) Then, since Approx(continuous) is greater than or equal to Approx(infinite), it is easy to see that: Thus, knowing the worst case error bound for the traveling salesan heuristic and the worst case difference between the optial values of the objective functions for the two probles, worst case error bounds can be found. 5. Coputational Experience Attepts to find error bounds for the worst case of the proble were not successful. Therefore, considerable coputational experientation was conducted to test the validity of the above approxiation for the case with infinite storage space. This case was selected for experientation because it reflects the other extree of the continuous processing case. For this purpose, the proposed approxiate algorith was prograed in FORTRAN to solve 600 probles ranging fro 3 to 7 jobs and 2 to 7 achines with the objective of iniizing akespan. Further, initial setup tie at each achine was assued to be sequence independent. The initial achine setup ties and the processing ties of these probles were generated fro a unifor distribution in the range (1,99). The sequence dependent setup ties of all jobs were generated frd a unifor distribution in the range (1,9). Each of these

9 N. D. Gupta probles was solved by using the proposed approxiate algorith described in Section 4 above. The traveling salesan proble representing the continuous processing flowshop proble in Section 4, represented by Equation (7), was solved to optiality using an existing search algorith [7]. The percent deviation of the approxiate akespan thus found fro the optiu akespan, found through coplete ent~eration, was coputed. Based on ZO probles of each size, the following statistics were collected: n n~ber of ties approxiate akespan was optial. 1 n Z nuber of ties the deviation of approxiate akespan fro the optial akespan was greater than zero but less than or equal to 3 percent. n3 nuber of ties the deviation of approxiate akespan fro optial akespan was greater than 3 percent but less than or equal to 5 percent. MIN: iniu percent deviation of the akespan fro optial akespan. AVR: average percent deviation of the approxiate akespan fro optial akespan. MAX: axiu percent deviation of the approxiate akespan fro optial akespan. Table 1 depicts these suary statistics for the 30 sets of ZO probles each. Fro Table 1, it is clear that the proposed algorith finds near optiu solutions in any cases, with percent deviation of akespan fro optiu akespan often being less than 5 percent. Therefore, it ay be concluded that the proposed approxlate algorith is quite effective in iniizing the akespan for the flowshop scheduling proble with infinite storage space in the presence of sequence dependent setup ties. To test the difference between the optial akespan of the continuous and the infinite storage cases, the percent deviation of the optial akespan for the continuous case fro the infinite storage space case optial akespan was coputed. Table Z shows these results where n, nz' n3 etc. need to be interpreted in ters of the percent difference between two optial 1 akespans.

10 Sequence-Dependent Flowshop Scheduling 215 Table 1 Percentage Deviation fro Optiu Makespans n M n 1 n 2 n3 AVR MIN MAX A coparison of Tables 1 and 2 shows that the optial schedule for the continuous case yields solutions of better quality than the differences in the two optial akespans. In other words, if e =0, 2 actual error is less than e where e is given in 3 l Table 2.

11 216 J. N. D. Gupta Table 2 Percent Deviation of Optial Makespan for Continuous and Infinite Storage Flowshops n M n 1 n 2 n3 AVR MIN MAX The coputational tie for the proposed approxiate algorith was not easured since it depends on the solution ethod used to solve the traveling salesan proble. The transforation of the flowshop scheduling proble to a traveling salesan probledl can be accoplished in polynoially bounded co-

12 Sequence-Dependent Flowshop Scheduling 217 putational effort. Therefore, if the approxiate algorith for the traveling salesan proble is polynoially bounded, so will the proposed approxiate algorith for the flowshop proble. 6. Conclusions This paper has discussed the flowshop scheduling proble with sequence dependent setup ties and shown that the proble is NP-coplete for the continuous processing and the liited or infinite or storage space available cases. A traveling salesan forulation of the continuous processing case has been proposed and used to find approxiate solutions to the liited or infinite storage space cases as well. Coputational results indicated that the proposed approxiate algorith was relatively effective in finding the optial or near optial solutions. While the finite storage space case was not considered during the coputational phase, it is felt that the proposed approxiate algorith should perfor better for this case because of its being soewhere in the iddle of the two cases considered. Further, soe coputational siplifications, like those indicated by Szwarc (19] for the sequence independent case can be used to decrease coputational effort of the proposed algorith. Acknowledgeents The author appreciates thoughtful eoents of two anonyous referees which iproved the clarity of this paper. References (1) Baker, K. R.: Introduction to sequencing and Scheduling, John Wiley and Sons, Inc., New York, N. Y. (1974). (2] Bellan, R., A. O. Esogbue and I. Nabeshia: Matheatical Aspects ot Scheduling and Applications, Pergaon Press, In., New York, N. Y. (1982). [3) Bonney, M. C., and S. W. Gundry: "Solutions to the Constrained Flowshop Sequencing Proble", Operational Research Quarterly, Vo1.27, No.4, pp (1976). (4] Corwin, B. D., and A. O. Esogbue: "Two-Machine Flowshop Scheduling Probles with Sequence Dependent Setup Ties: A Dynaic Prograing Approach," Naval Research Logistics Quarterly, Vol.21, pp (1974).

13 218 J. N. D. Gupta [5] Garey, M. R., and D. S. Johnson: Coputers and Intractability: A Guide to the Theory of NP-Copleteness, W. H. Freean and Copany, San Francisco, CA (1979). [6] Gupta, J. N. D.: "A Review of Flowshop Scheduling Research," in L. P. Ritzan et al (Eds) Disaggregation: Probles in Manufacturing and Service Organizations, Mart inus Nijhoff, The Hague (1979). [7] Gupta, J. N. D.: "A Search Algorith for the Traveling Salesan Proble", Coputers and Operations Research, Vol.5, No.4, pp (1978). [8] Gupta, J. N. D.: "Optial Flowshop Scheduling with no Interediate Storage Space", Naval Research Logistics Quarterly, Vol.27, No.2, pp (1976). [9] Gupta, J. N. D.: "A Search Algorith for the Generalized Flowshop Scheduling Proble", Coputers and Operations Research, Vol.2, No.2, pp (1975). [10] Gupta, J. N. D., and W. P. Darrow: "Two-Machine Flowshop Scheduling Probles with Sequence Dependent Setup Ties," Paper presented at the ORSA/TIMS Meeting, Orlando Florida, Noveber 7-9 (1983). [11] Gupta, J. N. D., andw. P. Darrow: "Approxiate Schedules for the Two Machine Flowshop with Sequence Dependent Setup Ties", Indian Journal of Manageent and Systes, Vol.l, No.l, pp.6-11 (1985). [12] Hefetz, N., and 1. Adiri: "The Influence of Missing Operations on Scheduling Probles", Naval Research Logistics Quarterly, Vo1.29, No.3, pp (1982). [13] Johnson, S. M.: "Optial Two- and Three-stage Production Schedules with Setup Ties Included," Naval Research Logistics Quarterly, Vol.l, pp (1954). [14] Levner, E. M.: "Optial Planning of Parts I Machining on a Nuber of Machines", Autoation and Reote Control, No.12, pp (1969). [15] Panwalkar, S. S., M. L. Sith and C. R. Woolla: "Counterexaples for Certain Flowshop Probles", Naval Research Logistics Quarterly, Vo1.28, No.2, pp (1981). [16] Parker, R. G., and R. L. Reardon: "The Traveling Salesan Proble: An Update of Research", Naval Research Logistics Quarterly, Vo1.30, No.l, pp (1983) [17] Srikar, B. N., and S. Ghosh: "An MILP Model for the N-job-M-Stage Flowshop with Sequence Dependent Setup Ties", Working Paper No.WPS 84-32, The ohio State University, Colurnbus, Ohio (1984). [18] Srikar, B. N., and S. Ghosh: "An Optial Algorith to Miniize Makespan in Sequence Dependent Flowshop", Working Paper No.WPS 84-57, The Ohio

14 Sequence-Dependent Flowshop Scheduling 219 State University, Colubus, Ohio (1984). [19] Szwarc, W.: "A Note on the Flowshop Proble without Interruptions in Job Processing," Naval Research Logistics Quarterly, Vo1.28, No.4, pp (1981). [20] Yoshida, T., and K. Hitoi: "Optial Two-Stage Production Scheduling with Setup Ties Separated," AIIE Transactions, vol.ll, pp.261-2b3 (1979). Jatinder N. D. GUPTA: Departent of Manageent Science, Ball State University, Muncie, Indiana 47306